Claude Code: Automatic Mode Secures AI Coding

A New Mode to Secure Coding with Claude Code

Anthropic recently unveiled a significant advancement for its language model Claude Code: an automatic mode. This mode is designed to allow the AI to manage permissions autonomously while integrating robust security measures. The goal is to provide a more secure alternative to the "dangerously skip permissions" option, often used by developers during intensive coding sessions.

Claude Code does not just generate code; it is also capable of executing essential shell commands for producing results. This includes tasks such as creating directories, moving files, or updating on GitHub. However, allowing an AI to operate unsupervised on a system can be concerning. To avoid potential disasters, Claude Code has implemented various permission systems.

Protections to Prevent Disasters

One of the security measures involves restricting Claude's operations to a specific folder hierarchy, which limits the risk of large-scale damage. However, this protection is not foolproof and does not prevent the possibility of Claude harming an entire codebase. Another security strategy involves requesting permissions for each potentially problematic action, which can slow down the development process. Instead of allowing Claude to write code fluidly, each command must be validated individually, which can be tedious.

Claude Code offers several levels of permissions, allowing users to choose the one that best fits their needs. Among these options, the "nuclear" option allows bypassing permission checks, although it carries significant risks.

Automatic Mode: A Compromise Between Security and Efficiency

Claude Code's automatic mode aims to minimize interruptions while reducing risks. Currently in testing phase, it is accessible only to users on the Team plan. Anthropic plans to extend it to users on the Enterprise and API plans in the near future.

According to Anthropic, "the automatic mode offers a middle ground, allowing for longer tasks to be executed with fewer interruptions while introducing less risk than skipping all permissions." Before each tool call, a classifier evaluates actions to identify those that could be destructive, such as mass file deletions or executing malicious code.

Actions deemed safe by the classifier are executed automatically, while those considered risky are blocked, prompting Claude to take a different approach. If Claude continues to attempt actions that are repeatedly blocked, a permission request is ultimately sent to the user.

The Limits and Precautions of Automatic Mode

Although the classifier in automatic mode is designed to identify potentially dangerous commands, it cannot eliminate all risks. It is highly recommended to work in isolated environments. As with any AI technology, automatic mode can sometimes be confusing. Some risky actions may be allowed if the AI does not correctly understand the context, while harmless actions may be blocked.

Currently, I cannot test this feature, as I am subscribed to the Max plan at $100/month, which does not offer this access. However, I have often been frustrated by the need to validate each permission when I simply want Claude to do its job. I regularly back up my system, so if automatic mode or the "dangerously skip permissions" option caused damage, I could restore my data.

I would likely be more inclined to use automatic mode rather than the completely open option. I would appreciate the increase in productivity while benefiting from safeguards. Once this feature becomes available for my plan, I look forward to trying it out.

Towards Controlled Autonomy in Coding

Currently, automatic mode is compatible with the Sonnet 4.6 and Opus 4.6 models. Anthropic specifies that "automatic mode may have a slight impact on token consumption, cost, and latency of tool calls." This approach is expected to improve over time, as Claude Code, although young, has already made a significant impact on the coding world.